Getty Images v Stability AI: What the High Court’s Decision Means for Rights-Holders and AI Developers

On 4 November 2025, the High Court published its much-anticipated judgment in Getty Images v Stability AI. This is the first major UK ruling to examine how copyright and trade mark laws apply to the development and deployment of generative AI models. Although Getty Images brought both primary and secondary copyright infringement claims and trade mark infringement claims, the Court ultimately rejected the central copyright allegation and found only limited liability for trade mark infringement. Nonetheless, the decision provides important guidance for rights-owners and AI developers alike, particularly on territoriality, the provenance of training datasets and the handling of outputs that may replicate marks or watermarks.

The Parties

Getty Images is a global provider of licensed photographic content, relying heavily on contributor agreements, exclusive licensing terms and brand recognition (including the well-known Getty and iStock watermarks).

Stability AI is a developer of generative AI models, including Stable Diffusion, an image-generation model partly trained on large-scale, publicly available datasets. Stability provides model weights (numerical parameters that determine the importance of features in a dataset) as open-source downloads and also operates hosted environments such as DreamStudio.

Relevant Factual Background

Getty discovered that some of its images seemed to appear in datasets used to train certain versions of Stable Diffusion. Getty alleged that Stability had trained its model using those images without consent, and that the resulting model could generate outputs replicating Getty’s trade marks and watermarks. Crucially, however, Getty accepted during proceedings that the training and development of Stable Diffusion did not occur in the United Kingdom. This had significant consequences for the viability of its copyright claims.

Stability acknowledged that some Getty images appeared in the datasets but denied infringement, arguing that an AI model’s learned weights are not copies of images and that any historical watermark reproduction was limited and mitigated by filtering processes.

The Claims

Getty brought several claims, the most significant being:

• Copyright infringement, including:
  •  A “Training & Development Claim” alleging primary infringement of Getty’s copyright through unlicensed training of Stable Diffusion’s AI model using Getty’s image library; and
  • Secondary copyright infringement under sections 22, 23 and 27 of the Copyright, Designs and Patents Act 1988 (“CDPA”), arguing that the model was an “infringing copy” imported or possessed in the United Kingdom.
• Trade mark infringement, focusing on the appearance of Getty/iStock watermarks in AI-generated outputs.
• Passing off, linked to similar concerns about watermarks and brand confusion.

The Court’s Analysis

1. The Training & Development Claim

Although this was the headline claim when proceedings began, Getty ultimately abandoned it because Getty could not demonstrate that the acts of reproduction or storage of Getty’s images occurred within the United Kingdom during the development of Stable Diffusion.

This procedural concession meant the Court did not consider or decide whether training an AI model on images protected by copyright, as a matter of law, constitutes infringement. Instead, the case centred on the secondary dealing provisions.

The key takeaway is that territoriality matters profoundly: if training of AI systems occurs outside the United Kingdom, a UK court may not consider primary copyright infringement claims relating to that training.

2. Secondary Copyright Infringement (Sections 22, 23 and 27 of CDPA)

This became the central legal issue. Getty argued that Stable Diffusion effectively contains copies of Getty’s images, and that making it available for download in the United Kingdom constituted importation of infringing copies, and as a result, infringed its copyright.

Stability contended that Stable Diffusion did not store or reproduce visual information from training images, and so cannot be infringing copies. Stability also denied knowledge of infringement.

Getty asserted that an “infringing copy” does not require that the imported article actually reproduces the copyright work, and the mere fact that its existence would have involved infringing copyright was enough.

The Court rejected this argument.

The Court held that AI model weights are not a “copy” of the images in the sense required by the CDPA. The CDPA requires a recognisable reproduction of a work or a derivative in which it is embodied. An AI model, by contrast, contains statistically trained parameters, not stored copies or reconstructions of photographs. Additionally, the model weights of the various iterations of Stable Diffusion did not store the visual information in the copyright works.

Accordingly:

• The Court ruled that the Stable Diffusion model is not an “infringing copy” for the purposes of section 27.
• Importation, possession or distribution of the model in the UK did not constitute secondary copyright infringement.

3. Trade mark Infringement

The Court was more sympathetic to Getty’s trade mark allegations. It accepted that in certain earlier versions of Stable Diffusion, and under realistic, non-contrived prompting, the model could generate synthetic images containing the Getty or iStock watermark.

If such outputs are generated “in the course of trade” and are capable of causing consumer confusion, they may constitute trade mark infringement.

However, the Court noted:

• The risk of watermark reproduction was not uniform across model versions;
• Filtering measures and improvements in hosted environments (e.g., DreamStudio) materially reduced the likelihood of trade mark-bearing outputs; and
• Watermark replication appeared less likely when users accessed the model via controlled platforms than through early open-source downloads.

Therefore, the Court concluded that there were some “limited” and “historic” instances of trade mark infringement from earlier iterations of Stable Diffusion. However, the Court found the infringement did not amount to detriment to distinctive character, detriment to reputation, or unfair advantage.

4. Passing Off

The Court declined to address the passing off claims in light of the trade mark infringement decision.

Key Takeaways

This is the first major UK ruling to examine how copyright and trade mark laws apply to the development and deployment of generative AI models. It is a significant development, albeit one that is likely to be appealed. Notwithstanding that, the following should be noted, assuming the decision withstands any future appeals:

For Rights-Holders and Content Providers

• If infringing training occurs outside the United Kingdom, the ability to bring UK copyright claims may be significantly reduced. Rights-holders should therefore monitor not only dataset usage but also where training activities are conducted.
• This case confirms that watermark replication, if it occurs in normal model use, can constitute a trade mark infringement claim. Regular monitoring of AI outputs for watermark misuse should become standard practice, and content owners should watermark their content.
• Clear contractual terms prohibiting or controlling AI training remain the most dependable mechanism for protecting content portfolios.

For AI Developers and Deployers

• The Court provided comfort by rejecting the argument that model parameters in themselves are infringing. However, this does not absolve developers of responsibility for outputs.
• The Court gave weight to Stability’s filtering improvements. Developers should maintain and document watermark detection, filtering pipelines, and safe-output mechanisms, especially in hosted environments.
• The factual analysis in the judgment underlines the value of clear internal documentation. Developers should maintain auditable records showing where data came from, what filtering was undertaken, and when.
• Although training abroad does not guarantee immunity, it can materially change the legal exposure within the United Kingdom. Developers should understand and record where training, storage and hosting activities occur.